Download Annotation Query Language

Text Extraction from Big Data
Дмитрий Брюхов, к.т.н., с.н.с.
Институт Проблем Информатики РАН
2
Content
 Text extraction in big data
 Span Algebra
 Annotation Query Language
 The Eclipse IDE for text analytics
 Example of creating text analytics project in Eclipse
3
What is big data?
 Massive volumes of data stored across many machines
 Too big to back up
 Could be structured, semi-structured or unstructured data
 Using traditional OLTP or data warehouse solutions for data
analysis are inadequate
4
What is Hadoop?
 Open source project
 Written in Java
 Optimized to handle
 Massive amounts of data through parallelism
 A variety of data (structured, unstructured, semi-structured)
 Using inexpensive commodity hardware
 Great performance
 Reliability provided through replication
 Not for OLTP, not for OLAP/DSS, good for Big Data
5
Terminology review
6
Hadoop architecture
 Two main components:
 Hadoop Distributed File System (HDFS)
 MapReduce Engine
7
Hadoop distributed file system (HDFS)
 Hadoop file system that runs on top of existing file system
 Designed to handle very large files with streaming data access
patterns
 Uses blocks to store a file or parts of a file
8
Map Reduce
9
Know where to look for the data
 Warehouse data about purchase history of a customer
 Site browsing history of the customer
 Comments on Websites
 Customer surveys
10
Integrate from various sources
 Different Data Source Platforms
 Different hardware platforms
 Different database systems
 Different Data Formats
 Structured Data Sources
 Semi Structured Data Sources
 Unstructured Data Sources
11
Problem with unstructured data
 Structured data has
 Known attribute types
 Integer
 Character
 Decimal
 Known usage
 Represents salary versus zip code
 Unstructured data has no
 Known attribute types nor usage
 Usage is based upon context
 Tom Brown has brown eyes
 A computer program has to be able view a word in context to
know its meaning
12
Need to harvest unstructured data
 Most data is unstructured
 Most data used for personal communication is unstructured
 Email
 Instant messages
 Tweets
 Blogs
 Forums
 Opinions are expressed when people communicate
 Beneficial for marketing
 Give insight between you and your competitors
13
Need for structured data
 Business intelligence tools work with structured data
 OLAP
 Data mining
 To use unstructured data with business intelligence tools
 Requires that structured data to be extracted from unstructured
and semi-structured data
 InfoSphere BigInsights provides a language, Annotation Query
Language (AQL)
 Syntax is similar to that of Structured Query Language (SQL)
 Builds extractors to extract structured data from
 Unstructured data
 Semi-structured data
14
Content
 Text extraction in big data
 Text Analytics in BigInsights
 Span Algebra
 Annotation Query Language
 The Eclipse IDE for text analytics
 Example of creating text analytics project in Eclipse
15
InfoSphere BigInsights
 InfoSphere BigInsights – это программная платформа,
призванная помочь организациям находить и анализировать
бизнес-информацию, скрытую в больших объемах
разнообразных данных, которые часто игнорируются или
отвергаются, поскольку слишком неудобны или трудны для
обработки традиционными средствами.
 Примерами таких данных являются данные социальных
медиа, новостные ленты, журналы событий, данные о
поведении потребителей (clickstream-данные), выходные
данные электронных датчиков и даже некоторые
традиционные транзакционные данные.
16
InfoSphereBiglnsights–A Full HadoopStack (IBM)
17
Key aspects of text analytics in BigInsights
 A declarative language for identifying and extracting content
from text data— The Annotation Query Language (AQL)
enables programmers to create views (collections of records)
that match specified rules.
 User-created or domain-specific dictionaries— Dictionaries can
identify relevant context across input text to extract business
insight from documents.
 User-created rules for text extraction— Pattern discovery and
regular expression (regex) building tools enable programmers
to specify how text should be analyzed to isolate data of
interest.
 Provenance tracking and visualization— Text analysis is often
iterative by nature, requiring rules (and dictionaries) to be built
upon one another and refined over time.
18
Text Analytics Overview
Jaql
AOG
Text
Analytics
AQL
Eclipse
Extractors
"Acquisition"
"Address"
"Alliance"
"AnalystEarningsEstimate"
"City"
"CompanyEarningsAnnouncement"
"CompanyEarningsGuidance"
"Continent"
"Country"
"County"
"DateTime"
"EmailAddress"
"JointVenture"
"Location" "Merger"
"NotesEmailAddress"
"Organization" "Person"
"PhoneNumber" "StateOrProvince"
"URL" "ZipCode"
Analyst
19
Content
 Text extraction in big data
 Text Analytics in BigInsights
 Span Algebra
 Annotation Query Language
 The Eclipse IDE for text analytics
 Example of creating text analytics project in Eclipse
20
Data and Execution Model
 Algebra operates over a simple relational data model with
three data types: span, tuple, and relation.
 A span is an ordered pair <begin, end> that denotes the region
of text
 A tuple is a finite sequence of w spans <s1, ..., sw>;
 we call w the width of the tuple.
 A relation is a multiset of tuples with the constraint that every
tuple must be of the same width.
 Each operator in our algebra takes zero or more relations as
input and produces a single relation as output.
21
Spans vs Strings
Moscow University from Moscow
 P<1,7> => “Moscow”
 P<24,30> => “Moscow”
 P<1,7> ≠ P<24,30>
22
Algebra Operators
 Relational Operators
 Span Extraction Operators
 Span Aggregation Operators
23
Relational Operators
 Since our data model is a minimal extension to the relational
model, all of the standard relational operators apply without
any change.
 Select
 Project
 Join
 …
 The main addition is that we use a few new selection
predicates applicable only to spans
 FollowsSpan
24
Span Extraction Operators
 A span extraction operator identifies segments of text that
match a particular input pattern and produces spans
corresponding to each such text segment.
 Regular expression matcher (Ere)
 Given a regular expression r, Ere(r) identifies all non-overlapping
matches when r is evaluated from left to right over the text
represented by s. The output of Ere(r) is the set of spans
corresponding to these matches.
 Dictionary matcher (Ed)
 Given a dictionary dict consisting of a set of words/phrases, the
dictionary matcher Ed(dict) produces an output span for each
occurrence of some entry in dict within doctext.
25
Span Aggregation Operators
 Span aggregation operators take in a set of input spans and
produce a set of output spans by performing certain
aggregate operations over their entire input.
 Block
 The block operator is used to identify regions of text where input
spans occur with enough regularity.
 Consolidate
 The block operator is used to identify regions of text where input
spans occur with enough regularity.
 Containment consolidation
 discard annotation spans that are wholly contained within other
annotation spans.
 Overlap consolidation
 produce new spans by merging overlapping spans
26
Span Aggregation Example
−− Define a dictionary of instrument names
create dictionary Instrument as ( ’ flute ’ , ’ guitar ’ , ... );
−− Use a regular expression to find names of band members
create view BandMember as
extract regex /[A−Z]\w+(\s+[A−Z]\w+)/
on 1 to 3 tokens of D.text
as name
from Document D;
−− A single ReviewInstance rule . Finds instances of BandMember followed
within 30 characters by an instrument name.
create view ReviewInstance as
select CombineSpans(B.name, I.inst) as instance
from
BandMember B,
(extract dictionary ’ Instrument’ on D.text as inst from Document D) I
where
Follows(B.name, I . inst , 0, 30)
consolidate on CombineSpans(B.name, I.inst);
27
Content
 Text extraction in big data
 Text Analytics in BigInsights
 Span Algebra
 Annotation Query Language
 The Eclipse IDE for text analytics
 Example of creating text analytics project in Eclipse
28
Annotation Query Language (AQL) overview
 Annotation Query Language (AQL) is the language for
developing text analytics extractors in the InfoSphere®
BigInsights™ Text Analytics system. An extractor is a program
written in AQL that extracts structured information from
unstructured or semistructured text.
 AQL is a declarative language, with a syntax that is similar to
that of the Structured Query Language (SQL).
29
AQL Data model
 Similar to the standard relational model
 Data is stored in tuples
 Data records of one or more columns or fields
 A collection of tuples forms a relation
 All tuples in a relation must have the same schema
 That is names and field types must be the same
 Scalar types
 Integer – 32-bit signed integer
 Float – Single precision floating-point number
 Text – Unicode string
 Span – Contiguous region of characters in a text object
 List – Represents a bag of values of type (Integer, Float, Text, or Span)
30
AQL Execution model
 An AQL extractor consists of a collection of views
 Each view defines a relation
 Special view called Document
 Represents the document that is being annotated
 Two fields in this view
 text –Textual content of the document
 label –Label for the document, usually the name of the document
 Support for include statements
 Allows for reused of code
31
AQL components overview
 Create view statement
 Creates a view name and defines the tuples inside the view
 Extract statement
 Provides functionality for extracting features directly from text
 Select statement
 Mechanism for constructing complex patterns out of simpler
building blocks
 Detag statement
 Pre-processing step to remove all HTML and XML tags from a
document
32
AQL components overview (cont.)
 Create dictionary
 Define dictionaries of words or phrases to use in extract statements
 Create table
 Similar to the SQL create table statement
 Built-in functions
 Functions for use in extraction rules
 User-defined functions
 Allows user to define customer functions to be used in extraction
rules
33
Create View Statement
 Creates a view and defines the tuples inside that view.
create view <viewname> as <select or extract statement>;
 The output view statement defines a view to be an output
view.
output view <viewname>;
 Example
create view PhoneNumbers as
extract regex /(\d{3})(-)(\d{3})(-)(\d{4})/ on d.text return
group 1 as areacode and
group 3 as exchange and
group 5 as extension
from Document d;
34
Regular expressions
 Regular expressions are used to extract textual values from a
non-structured file
 AQL uses Perl syntax for regular expressions . The format is:
regex[es] /<regex1>/
[and /<regex2>/ and ... and /<regex n>/]
[with flags '<flags string>']
on [<token spec>] <name>.<column>
<grouping spec>
 Grouping specification determines how to handle capturing
groups in the regular expression. Capturing groups are regions
of the regular expression match, identified by parentheses in
the original expression.
 Example
extract regex
on d.text
group
group
group
/(\d{3})(-)(\d{3})(-)(\d{4})/
return
1 as areacode and
3 as exchange and
5 as extension
35
Dictionaries
 Use the dictionary extraction specification to extract strings
from input text that are contained in a dictionary of strings.
 Dictionaries are created either using the create dictionary
stateme
 Dictionary extraction specification
dictionar[y|ies]
'<dictionary>‘
[and '<dictionary>' and ... and '<dictionary>']
[with flags '<flags string>']
 Example
create dictionary ConjunctionDict as
('and', 'or', 'but', 'yet');
create view Conjunction as
extract
dictionary 'ConjunctionDict‘ on D.text as name
from Document D;
36
Splits
 Use the split extraction specification to split a large span into
several smaller spans.
 The split extraction specification takes two arguments:
 A column that contains longer target spans of text
 A column that contains split points
 Example
create view Sentences as
extract
split using B.boundary
retain right split point
on B.text
as sentence
from (
extract
D.text as text,
regex /(([\.\?!]+\s)|(\n\s*\n))/ on D.text as boundary
from Document D
) B;
37
Blocks
 Use the blocks extraction specification to identify blocks of
contiguous spans across input text.
 The block extraction specification takes two parameters:
 count - Specifies how many spans can make up a block
 separation - Specifies how much distance is allowed between
spans before they are no longer considered to be contiguous
 Example
create view BlockCapitalWords as
extract blocks
with count between 2 and 3
and separation between 0 and 100 characters
on cw.words as captialwords
from CapitalizedWords cw;
38
Part of speech
 Use the Part-of-speech extraction specification to identify
locations of different parts of speech across the input text.
 Syntax:
part_of_speech
'<part of speech spec>'
[and '<part of speech spec>']*
[with language '<language code>']
[and mapping from <mapping table name>]
on <input column> as <output column>
from <input view>
 part of speech spec
 a comma-delimited list of part-of-speech tags
 A combination of an internal part-of-speech name and flags, as
defined by the mapping table
 Example
create view Nouns as
extract part_of_speech 'NN','NNS'
on D.text as noun
from Document D;
39
Multilingual support
Language
Afrikaans
Arabic
Catalan
Chinese (Simplified)
Chinese (Traditional)
Czech
Danish
Dutch
English
Finnish
French
German
Greek
Italian
Japanese
Korean
Norwegian (Bokmal)
Norwegian (Nynorsk)
Polish
Portuguese
Russian
Spanish
Swedish
Code
af
ar
ca
zh-CN
zh-TW
cs
da
nl
en
fi
fr
de
el
it
ja
ko
nb
nn
pl
pt
ru
es
sv
Tokenization
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Part-of-speech analysis
N
Y
N
Y
Y
N
Y
Y
Y
N
Y
Y
N
Y
Y
N
N
N
N
Y
N
Y
N
40
Sequence patterns
 Use the pattern extraction specification to perform pattern
matching across an input document in addition to spans
previously extracted from the input document.
 Allows you to write complex extraction patterns in a compact
fashion involving
 Alternation
 Sequences
 Repetitions
 Example
create view CountryLocation as
extract
pattern (<C.country>) <Token>{5,35} (<L.longlat>)
return group 1 as country
and group 2 as location
from Countries C, Locations L;
41
The select statement
 Allows for the construction of more complex patterns from
simpler building blocks
 The basic for is very similarly structured to an SQL select
statement
 Format
select <selectlist>
from <fromlist>
[where <predicate>]
[consolidate on <column>]
[group by <group list>]
[order by <order list>]
[limit <maximum number of tuples for each document>]
42
AQL built-in functions
 AQL has three types of built-in functions
 Predicate
 Scalar
 Aggregate
 Predicate functions- Building blocks for the where clause
 Contains, Equals, Follows
 Scalar functions - Return integers, strings or new spans
 CombineSpans, GetBegin, GetEnd, GetText
 Aggregate functions - Create a single value from a list of input
values
 Avg, Count, Max
43
User-defined functions
 AQL allows user to define custom functions used in extraction
rules
 Only scalar user-defined functions (UDFs) are supported
 Currently AQL supports UDFs implemented in Java. Implement
the UDF as a public method in a Java class.
 You make the user-defined functions (UDFs) available to AQL
by using the create function statement.
 Scalar user-defined functions (UDFs) can be used in the select
and where clauses in the same way as built-in scalar functions.
 User-defined functions that return Boolean can be used as
predicates in where and having clauses.
44
Pre-built extractor libraries
 The Standard and Multilingual tokenizer JAR files export the
following views:
Person
Organization
Location
Address
City
Continent
Country
County
DateTime
EmailAddress
NotesEmailAddress
PhoneNumber
StateOrProvince
Town
URL
ZipCode
CompanyEarningsAnnounce
ment
AnalystEarningsEstimate
CompanyEarningsGuidance
Alliance
Acquisition
JointVenture
Merger
45
Content
 Text extraction in big data
 Text Analytics in BigInsights
 Span Algebra
 Annotation Query Language
 The Eclipse IDE for text analytics
 Example of creating text analytics project in Eclipse
46
Developing text analytics extractors
 You can build and maintain extractors inside a full-fledged
Eclipse development environment by installing the InfoSphere
BigInsights tools for Eclipse.
 Inside the tools for Eclipse, there are two ways to develop an
extractor:
 Text Analytics Workflow perspective
 This perspective displays several panes, one of which is the Extraction
Tasks pane. This pane guides you through a six-step process to develop,
test, and export extractors for deployment. You start by selecting your
input document collection and conclude with exporting your working
extractor to the cluster.
 AQL Editor
 This editor provides a framework to develop and test AQL extractors.
47
The Eclipse IDE for text analytics
48
Eclipse extraction tasks and plan views
49
Label examples
50
Adding clues to the puzzle
51
Create AQL statements
52
Regular expression helpers
53
Regular Expression Generator
54
Regular Expression Builder
55
Testing your code
56
Extractor run results
57
Document tree view
58
Text Analytics Lifecycle
59
Content
 Text extraction in big data
 Text Analytics in BigInsights
 Span Algebra
 Annotation Query Language
 The Eclipse IDE for text analytics
 Example of creating text analytics project in Eclipse
60
Создание проекта BigInsights
61
Создание директории data
62
Импорт данных
63
Входной файл Facts.txt
"====================================================================
Afghanistan
Introduction
Afghanistan
Background: Afghanistan's recent history is characterized by war and
civil strife, with intermittent periods of relative calm and …
Geography Afghanistan
Location:
Southern Asia, north and west of Pakistan, east of Iran
Geographic coordinates:
Map references:
Area:
33 00 N, 65 00 E
Asia
total: 647,500 sq km water: 0 sq km land: 647,500 sq km
Area - comparative:
slightly smaller than Texas
64
Создание AQL файла
65
Создание представления Countries
module main;
create view Countries as
extract regex // on D.text
return group 1 as country
from Document D;
output view Countries;
66
Конструктор регулярных выражений
67
Добавление регулярного выражения в
представление Countries
Geography Afghanistan
module main;
create view Countries as
extract regex /Geography {1,}([A-Z][a-z]{1,} ?[A-Z]*[a-z]*)/ on
D.text
return group 1 as country
from Document D;
output view Countries;
68
Конфигурация выполнения
69
Результат выполнения экстракторов
70
Представление результата в виде дерева
документа
71
Извлечение координат
Geographic coordinates:
0 48 N, 176 38 W
create view Locations as
extract regex /Geographic coordinates: {1,}((\d{1,2} \d{2} [NS]),
(\d{1,3} \d{2} [EW]))/ on D.text
return group 1 as longlat
and group 2 as longitude
and group 3 as latitude
from Document D;
72
Работа со словарями
Exchange rates: afghanis per US dollar - 4,700 (January 2000), 4,750
(February 1999), 17,000 (December 1996), 7,000 (January 1995), 1,900
(January 1994), 1,019 (March 1993), 850 (1991); note - these rates reflect
the free market exchange rates rather than the official exchange rate,
which was fixed at 50.600 afghanis to the dollar until 1996, when it
rose to 2,262.65 per dollar, and finally became fixed again at 3,000.00
per dollar in April 1996
Fiscal year: 21 March - 20 March
create dictionary MonthsDict as
(
'January', 'February', 'March', 'April', 'May', 'June', 'July',
'August', 'September', 'October', 'November', 'December‘
);
create view Months as
extract
dictionary 'MonthsDict‘
on D.text as month
from Document D;
73
Работа с частями речи
Introduction
Afghanistan
Background: Afghanistan's recent history is characterized by war
and civil strife, with intermittent periods of relative calm and
stability. The Soviet Union invaded in 1979 but was forced to withdraw 10
create view Nouns as
extract part_of_speech 'NN' and 'NNS' with language 'en'
on D.text as noun
from Document D;
74
Работа с блоками
Introduction
Afghanistan
Background: Afghanistan's recent history is characterized by war
and civil strife, with intermittent periods of relative calm and
create view BlockNouns as
extract blocks
with count between 2 and 3
and separation between 0 and 50 characters
on n.noun
as BlockedNouns
from Nouns n;
75
Работа с паттернами
Geography Afghanistan
Location: Southern Asia, north and west of Pakistan, east of Iran
Geographic coordinates: 33 00 N, 65 00 E
create view CountryLocation as
extract
pattern (<C.country>) <Token>{5,35} (<L.longlat>)
return group 1 as country
and group 2 as location
from Countries C, Locations L;
output view CountryLocation;
76
Работа с Select
create view Textnames as
select GetText(C.country) as country, GetText(C.location) as location
from CountryLocation C;
77
Questions ?
78
Анонс
 Практикум
 27 марта, 18:30
 Машинный зал № 4